Yes it is data which isn't being used for training. One reason for doing this is patterns in data from a two years a go is less relevant then more recent data. Another reason is for cross validation each validation set must be different and to make each fold/backtest fairly comparable they should be trained on a equal amount of time (or same number of rows). To do this the available data is need for the later backtests to use.
Of course this is very use case dependent and it may make sense to use all the data.